AI Agents in Depth epistemic report

Type: agentic-system-analyses/types/agentic-system-epistemic-report.md

Source-and-claim boundary

The question is which selected teaching workflows acquire or produce truth-apt content, what checks license reliance, and which consumers give their results consequential force. The boundary governs the suite's complete artifact, partial loop. Evidence identity and layers are in its Source register; all added inspection uses SRC-1 at that same commit. No target execution was performed or adopted as observed evidence.

The assessed families are context-agent observations and replies, mode-dependent delivery, the ablation's numerical assessor and reporting consumer, and retry-policy candidate checking and disposition. Imported PDF assertions, financial results, and model answers are not interchangeable with executable retry rules. Claims found are the teaching purpose RT-CLM-context-purpose, bounded eligibility RT-CLM-canary-decision, and the tool's live-rate description EPI-CLM-current-rates. The assessor's stronger provenance wording is examined under EPI-CLM-grounding-source.

Entry point or operation	Source path	Coverage and records	Limit and conclusion prevented
CLI single, interactive and ablation dispatch	SRC-1: chapter1/context/main.py	Direct read of CLI dispatch, assessor call, warning and summary/table consumers; supplied RT-RTE-context-task, RT-RTE-context-ablation	No observed invocation or benefit
Four registered context tools and loop	SRC-1: chapter1/context/agent.py	Supplied complete agent implementation inspection and fixed dispatch; direct read of PDF, conversion, calculator and relevant tool descriptions, plus complete chapter1/context/calc_sandbox.py; EPI-RTE-pdf-import, EPI-RTE-currency-conversion, RT-RTE-calculation-path, RT-RTE-code-execution	Arbitrary interpreter programs not inspected; their semantic warrant cannot be generalized
Numerical extraction and grounding verdict	SRC-1: chapter1/context/grounding.py	Complete direct read; EPI-RTE-grounding-check, EPI-RTE-grounding-report	No reference task rubric; no derivation-provenance oracle
Context omission, reset and task continuation	SRC-1: chapter1/context/agent.py, chapter1/context/main.py	Supplied RT-RTE-context-ablation, RT-OBJ-conversation-history	Retention does not guarantee delivery or model compliance
Candidate demo, campaign and exposed release function	SRC-1: chapter9/self-modifying-agent/demo.py, chapter9/self-modifying-agent/run_experiment_9_6.py, chapter9/self-modifying-agent/evolution.py	Supplied RT-RTE-policy-candidate; direct read of generation control, validation and manifest expression in evolution	No deployed successor or live rollback; campaign report writer remains uninspected
Container validation and metrics	SRC-1: chapter9/self-modifying-agent/candidate_sandbox.py, chapter9/self-modifying-agent/sandbox_runner.py	Supplied RT-RTE-candidate-validation, RT-BAP-candidate-policy	No inspected execution, Docker image verification or adversarial oracle-integrity proof
Provider resolution and model state	SRC-1: agentbook/providers/resolution.py, chapter9/self-modifying-agent/llm_generator.py	Supplied RT-CMP-context-model, RT-CMP-coding-model	Weights and provider internals unavailable; immutable versions and parameter adaptation not established
User-memory maintenance and later consumption	SRC-1: chapter3/user-memory/	Initial boundary witnesses only; targeted symbol search in memory_manager.py and background_memory_processor.py is not a full route inspection	Add/update/delete, consolidation, conflict handling, admission and prompt reliance remain unassessed here; prevents a suite-complete negative about acceptance or integration
Other chapters, instruction-loading skills, training, retrieval, collaboration and retained validation artifacts	SRC-1: chapter2/ through chapter10/, skills/, chapter-local validation paths	Available but uninspected except the selected chapter3 witnesses and chapter9 candidate workflow	No exhaustive architecture, evaluator, learning or knowledge-production conclusion
External reproduction repositories, weights, services and agent host	See boundary exclusions and dependencies	Not registered evidence	No end-to-end reproduction, host compliance or service-behavior conclusion

No additional hooks or dynamic tool registrations are established in the selected context dispatcher. That is a bounded supplied finding, not an absence claim for the suite. Available retained artifacts were not used to infer which routes ran.

Epistemic-object inventory

Object ID	Candidate truth-apt content or none	Claimed role	Evidence source ID and local anchor	Gap/limit
EPI-OBJ-pdf-result	External PDF assertions, as extracted text	see record	SRC-1: chapter1/context/agent.py	Author truth and extraction fidelity unchecked
EPI-OBJ-currency-result	Conversion under fixed rates and rounded numerical result	see record	SRC-1: chapter1/context/agent.py	Financial freshness unknown; timestamp is calculation time
EPI-OBJ-final-answer	Task-dependent assertions when reply has them; a refusal need not	see record	SRC-1: chapter1/context/agent.py, chapter1/context/grounding.py	Proposition/domain requires a concrete reply; source contains no observed reply adopted here
RT-OBJ-conversation-history	Mixed container: tasks, guidance, assistant and tool text	see record	SRC-1: chapter1/context/agent.py	Not one homogeneous epistemic object; the three parts above separate material content
RT-OBJ-retry-policy	Formal retry/circuit policy; comments and impact predictions assert intended behavior	see record	SRC-1: chapter9/self-modifying-agent/evolution.py	Policy instructions themselves are behavioral rules, not established claims of improved real-world performance
RT-CMP-context-model	Internal truth-apt units not determinable	see record	SRC-1: agentbook/providers/resolution.py	Opaque parameters, mutable routing, no weight-level inspection
RT-CMP-coding-model	Internal truth-apt units not determinable; external proposal is module text	see record	SRC-1: chapter9/self-modifying-agent/llm_generator.py	Proposal request does not establish compliance or weight learning

Authority-route ledger

Each row is one function. A repeated route citation separates functions of an existing implementation without reallocating its identity. Fields owned by a shared record say see; these are not fresh declarations. All implemented statuses below denote source wiring, not observed operation.

Route ID	Route function	Architectural status	Object/candidate ID	Content/update relation	Transition or check target	Evaluator/condition and domain	Activation and timing	Possible or observed result	Implemented force	Epistemic authority and scope	Operational authority: behavior permitted, blocked, or changed	Behavioral-authority path: consumer, channel, force, horizon	Evidence source ID and local anchor	Claim IDs or none	Mismatch marker or none	Gap/limit
EPI-RTE-pdf-import	content transformation	implemented	see record	truth-apt transformation: acquisition/import	PDF bytes to text observations	see record	see record	see record	see record	Source warrant unknown; records source URL and page, not validation	Makes extracted assertions available for a later completion	see record	SRC-1: chapter1/context/agent.py	none	none	Parser success is not endorsement
EPI-RTE-currency-conversion	content transformation	implemented	see record	truth-apt transformation: entailed derivation	Numerical amount under fixed rate table	see record	see record	see record	see record	Conditional arithmetic only, assuming supplied amount and table; no live-rate warrant	Returned result can shape later reply	see record	SRC-1: chapter1/context/agent.py	EPI-CLM-current-rates	current versus static	Numeric input and rate truth not checked
RT-RTE-calculation-path	check/evidence production	implemented	see record	no content change	Expression AST admissibility	see record	see record	see record	see record	Syntax whitelist only, not source truth or physical units	Blocks prohibited AST before eval	see record	SRC-1: chapter1/context/calc_sandbox.py	none	none	Resource exhaustion and other tool routes outside check
RT-RTE-calculation-path	content transformation	implemented	see record	truth-apt transformation: entailed derivation	Evaluation of accepted mathematical expression	see record	see record	see record	see record	Result conditional on expression, Python arithmetic and supported names; no warrant for modeling assumptions	Produces tool observation	see record	SRC-1: chapter1/context/agent.py, chapter1/context/calc_sandbox.py	none	none	Caret rewrite changes Python syntax intentionally; input 2^3 becomes 2**3 and denotes 8, not Python XOR
RT-RTE-code-execution	content transformation	implemented	see record	truth-apt transformation: indeterminate	Arbitrary supplied Python output	see record	see record	see record	see record	Execution success licenses only that code completed on that call, not output truth	Output enters tool history	see record	SRC-1: chapter1/context/agent.py	none	none	A specific program is needed to decide acquisition, reshaping, derivation or conjecture; policies may be non-truth-apt
RT-RTE-context-task	content transformation	implemented	EPI-OBJ-final-answer	truth-apt transformation: indeterminate	Task/context to generated reply	see record	see record	see record	see record	Candidate generation only; no task correctness oracle	May finish the task and return text	see record	SRC-1: chapter1/context/agent.py	none	completion is not correctness	Concrete content needed to decide entailment versus non-entailment; terminal refusals need not produce claims
RT-RTE-context-task	retention	implemented	RT-OBJ-conversation-history	no content change	Append messages and trajectory	see record	see record	see record	see record	Retention grants no endorsement	Same live agent can expose earlier content to later tasks	see record	SRC-1: chapter1/context/agent.py, chapter1/context/main.py	none	none	Written single-task JSON has no inspected reload
RT-RTE-context-ablation	operational admission/selection/consumption	implemented	RT-OBJ-conversation-history	no content change	Stored versus delivered messages/tools/results	see record	see record	see record	see record	Visibility manipulation, not truth checking	Permits/withholds context at provider request	see record	SRC-1: chapter1/context/agent.py	RT-CLM-context-purpose	none	Hidden provider reasoning is not removed by dropping exposed reasoning text
EPI-RTE-grounding-check	check/evidence production	implemented	EPI-OBJ-final-answer	truth-apt transformation: entailed derivation	Extracted quantities versus task/tool-message quantities	see record	see record	see record	see record	Reports numeric-match predicate, not correctness or actual derivation lineage	No admission veto in this function	see record	SRC-1: chapter1/context/grounding.py	EPI-CLM-grounding-source	source availability overstatement	Ignores sub-floor answer figures; observations trigger abstention; valid arithmetic can yield unmatched numbers
EPI-RTE-grounding-report	operational admission/selection/consumption	implemented	see record	no content change	Grounding verdict to report fields	see record	see record	see record	see record	Advisory report only	Warns, labels tables, counts unsupported answers and assigns failure reason	see record	SRC-1: chapter1/context/main.py	RT-CLM-context-purpose, EPI-CLM-grounding-source	report label exceeds predicate	No answer withdrawal, rerun or policy replacement is implemented by this consumer
RT-RTE-policy-candidate	behavior/policy adaptation	implemented	RT-OBJ-retry-policy	non-truth-apt policy/content update: retry/circuit module proposal	Failure diagnosis and deterministic or model patch	see record	see record	see record	see record	Proposal, not established mechanism or transfer claim	Candidate source executes only during evaluation on inspected paths	see record	SRC-1: chapter9/self-modifying-agent/evolution.py, chapter9/self-modifying-agent/llm_generator.py	none	none	Diagnosis/predictions assert causal content; synthetic grouping and model proposal alone do not warrant that content
RT-RTE-candidate-validation	check/evidence production	implemented	RT-OBJ-retry-policy	truth-apt transformation: entailed derivation	Compilation, deny-list, signatures and retry/circuit outcomes on supplied cases	see record	see record	see record	see record	Conditional bounded software checks, given trustworthy runner/image and encoding; not general safety or improved capacity	False checks feed release rejection; metrics do not independently install changes	see record	SRC-1: chapter9/self-modifying-agent/evolution.py, chapter9/self-modifying-agent/sandbox_runner.py	RT-CLM-canary-decision	none	Candidate/evaluator share process; bounded tests do not authenticate causal diagnosis
RT-RTE-policy-candidate	disposition/acceptance	implemented	RT-OBJ-retry-policy	no content change	Changed candidate plus nonempty all-true supplied checks	see record	see record	see record	see record	Evidence-consuming eligibility against supplied checks, intended use shadow traffic with stable unchanged; no accepted general performance theory	Emits release_to_canary or reject_candidate manifest	see record	SRC-1: chapter9/self-modifying-agent/evolution.py	RT-CLM-canary-decision	eligibility is not activation	Direct callers can substitute check set; manifest expression does not enforce check-name completeness
RT-RTE-policy-candidate	retention	implemented	RT-OBJ-retry-policy	no content change	Save candidate source, checks, diagnosis and manifest	see record	see record	see record	see record	Candidate ledger, including rejected proposals; not acceptance by storage	Failed-control results can guide later coding request	see record	SRC-1: chapter9/self-modifying-agent/demo.py, chapter9/self-modifying-agent/run_experiment_9_6.py	none	none	Retention of accepted and rejected artifacts is not post-acceptance lifecycle integration
RT-RTE-policy-candidate	lifecycle integration	no route found within boundary	RT-OBJ-retry-policy	no content change	Accepted policy into stable operation	No installed-successor consumer in supplied selected paths	Not observed	No traced live installation	None established	No additional license from manifest	Stable remains unchanged in inspected demo/campaign	No live consumer established; evaluation authority remains separate	SRC-1: chapter9/self-modifying-agent/demo.py, chapter9/self-modifying-agent/evolution.py, chapter9/self-modifying-agent/run_experiment_9_6.py	RT-CLM-canary-decision	none	Bounded to selected candidate plane; other suite consumers unassessed

System-claim versus route comparison

Claim ID	Claimed operation or warrant	Claim source ID/anchor and evidence layer	Doctrine/design support	Implemented route IDs	Observed-run support	Causal support and design limits	Supported conclusion	Mismatch/unknown
RT-CLM-context-purpose	see record	SRC-1: chapter1/context/agent.py; doctrine/design	Ablation teaching intent	RT-RTE-context-ablation, EPI-RTE-grounding-check, EPI-RTE-grounding-report	None inspected	Context arms are wired; no retained intervention results inspected; completion and numeric matching do not identify task benefit	Implements a comparison/report mechanism	Importance/effect magnitude not established
EPI-CLM-current-rates	see record	SRC-1: chapter1/context/agent.py; doctrine/design in schema description	Description delivered to model	EPI-RTE-currency-conversion	None inspected	Constant table and calculation timestamp do not measure current exchange rates	Conditional conversion from shipped constants	Live-rate wording contradicted by implementation
EPI-CLM-grounding-source	see record	SRC-1: chapter1/context/grounding.py; doctrine/design	Explicitly weaker than correctness, but labels unmatched figures as without evidence	EPI-RTE-grounding-check, EPI-RTE-grounding-report	None inspected	Number matching omits arithmetic closure and actual provenance; observation presence forces abstention	A bounded reproducible text predicate	Does not establish that unmatched numbers were fabricated or lacked an evidential derivation
RT-CLM-canary-decision	see record	SRC-1: chapter9/self-modifying-agent/evolution.py; implementation and doctrine/design	Shadow-only eligibility, stable unchanged	RT-RTE-policy-candidate, RT-RTE-candidate-validation	None inspected	Fixed cases and expected booleans; direct caller check substitution; no live traffic	Candidate admission against caller-supplied evidence	Not production activation, live rollback or exercised self-improvement

Bounded conclusion

Imported assertions have unknown warrant. PDF parsing preserves a URL and page association, but neither authenticates the author nor checks extraction against the document. Delivery of that result can affect the completion model without a content acceptance decision. Context retention is separate from delivery, and delivery is separate from compliance. See EPI-RTE-pdf-import, RT-OBJ-conversation-history and RT-RTE-context-ablation.

Arithmetic produces conditional results, not source truth. The calculator evaluates accepted syntax; currency conversion derives a rounded result from a fixed table and supplied amount. Neither check warrants the task's assumptions, amount provenance, real-world units or rate freshness. Model-generated replies and arbitrary code outputs remain semantically indeterminate without concrete inputs/outputs; failure to prove entailment is not evidence of ampliative conjecture. See RT-RTE-calculation-path, EPI-RTE-currency-conversion and RT-RTE-code-execution.

The numerical assessor has advisory force, not a truth or provenance veto. Its target is extracted revenue-scale answer quantities compared with quantities available in task/tool text. It licenses the computed match/abstention classification, not actual derivation history. Its consumer warns and annotates reports; it does not reject the already returned answer or revise the agent. The no-observation arithmetic counterexample in EPI-RTE-grounding-check bounds the stronger evidence wording. No supersession of the supplied runtime is required: its statement that this implementation was uninspected there is valid, and this report supplies the missing local assessment.

Retry-policy acceptance is narrow but real at the manifest decision. The intended use and scope are shadow eligibility with stable unchanged. The ordinary validator consumes proposed code and expected software outcomes; release requires a changed candidate and all supplied checks true. A direct manifest caller can bypass the normal check-set construction. Thus acceptance is evidence-consuming in the normal driver, but not a universal release authorization. Candidates, including rejects, are retained. The rejected-control summary can shape a later coding request; that does not install a successor or accept the model's causal explanation. See RT-RTE-policy-candidate and RT-RTE-candidate-validation.

The diagnosis and policy module have localized content, and evaluation consumes the formal rule. The runtime separately supports implemented criticism and only afforded content-dependent iteration. Learning benefit, complete reflection, autonomous fulfillment of every theory-builder condition and exercised self-improvement remain independent unestablished properties, not a bundled negative. Missing installed live authority and subsequent causally dependent operation prevent attributing exercised self-improvement over this campaign horizon. The direct policy proposal/evaluation route still exists. See the theory and self-improvement account in RT-RTE-policy-candidate.

No accepted ampliative knowledge claim or post-acceptance stable integration is established in these selected routes. This is not a suite-wide absence: user-memory conflict handling, other evaluation/training mechanisms, external-host skills and available retained results remain unassessed. Provider parameter state and immutable answering-model identity also remain uninspected; neither inference calls nor symbolic candidate updates establish parametric learning. See RT-CMP-context-model, RT-CMP-coding-model and the coverage limits above.

Shared records

Operative objects

EPI-OBJ-pdf-result

Label: PDF tool observation

Part of: RT-OBJ-conversation-history

Source-native identity: parse_pdf result carried as tool content in history. Representational form: natural-language page text with symbolic URL/page/metadata envelope. Storage substrate: returned Python dictionary, serialized into in-process tool history and optional caller reports. It is the PDF observation part, not the mixed container or model answer. SRC-1: chapter1/context/agent.py; see EPI-RTE-pdf-import for supporting implementation passages and delivery limits.

EPI-OBJ-currency-result

Label: Currency tool observation

Part of: RT-OBJ-conversation-history

Source-native identity: convert_currency result carried as tool content. Representational form: symbolic numeric result dictionary, interpreted as a financial assertion by the completion model. Storage substrate: Python dictionary and serialized in-process tool history. Distinct from PDF assertions and free-form final replies, with an arithmetic consumer before model interpretation. SRC-1: chapter1/context/agent.py; evidence in EPI-RTE-currency-conversion.

EPI-OBJ-final-answer

Label: Terminal model reply

Part of: RT-OBJ-conversation-history

Source-native identity: terminal assistant message content selected as final_answer. Representational form: natural language, possibly containing numbers and formatting; storage substrate: live history, returned result and optional JSON report. A material part of the conversation container, not the entire container. Its claims and checks differ from imported observations. SRC-1: chapter1/context/agent.py; RT-RTE-context-task supplies termination evidence. The response can be a refusal; nonempty text establishes completion only. EPI-RTE-grounding-check supplies the separate numerical evaluator evidence.

Routes

EPI-RTE-pdf-import

Label: PDF import observation

Part of: RT-RTE-context-task

Endpoints/progression: local path or URL → bytes → pypdf.PdfReader → per-page extracted text → tool result → selected later provider request. Owner: ToolRegistry.parse_pdf and the context loop. Object: EPI-OBJ-pdf-result. URL and page structure preserve an association, not independent source truth. This is the PDF acquisition subroute of the supplied task route; it does not redefine its dispatcher or final-answer semantics. SRC-1: chapter1/context/agent.py, implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: URL, page count, per-page text and metadata, or an error dictionary.
• Later read-back: Same live agent can deliver retained tool messages on later task calls; reset/replacement discards them.
• Delegated visibility: Provider gets serialized tool content only when context mode permits; there is no separate extraction worker.
• Selection predicate: Model/direct caller supplies path or URL; protocol/path tests choose local file versus HTTP retrieval; page iteration extracts all parsed pages.
• Invalidation or expiry: No source revalidation or endorsement expiry; local history reset governs live retention.
• Activation or effect: Import and provider delivery are wired; changed model behavior unobserved. Consumer: completion model; channel: tool message; force: informational/advisory; horizon: later visible turns/tasks on the live agent.
• Evidence limits: No content evaluator, author-authentication check or extraction-fidelity comparison inspected here; source warrant unknown. Parsing success is not an acceptance decision.

text = page.extract_text() text_content.append({ "page": page_num, "text": text }) --- chapter1/context/agent.py

EPI-RTE-currency-conversion

Label: Currency conversion observation

Part of: RT-RTE-context-task

Endpoints/progression: amount/currency arguments → normalization → fixed-rate lookup → USD intermediate → rounded target amount → tool result → later model request. Owner: ToolRegistry.convert_currency and the loop. Object: EPI-OBJ-currency-result. Evaluator/condition: numeric parsing and membership in shipped table; arithmetic domain is conversion under those constants. This is the conversion subroute, not the whole supplied task loop. SRC-1: chapter1/context/agent.py, implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Original amount, normalized currencies, rounded converted amount/rate and calculation timestamp, or error.
• Later read-back: Retained tool observation can reach a later task invocation on the same live agent, subject to context mode.
• Delegated visibility: Provider receives serialized conversion result, not a live exchange-rate feed.
• Selection predicate: Dispatcher/direct call chooses conversion; unsupported currency returns error before arithmetic.
• Invalidation or expiry: No rate refresh or financial-validity expiry; history reset/replacement ends live retention.
• Activation or effect: Conditional calculation is wired; actual effect on answers unobserved. Consumer: completion model; channel: tool message; force: informational/advisory; horizon: subsequent visible turns/tasks.
• Evidence limits: Static rate premises lack independently inspected financial warrant. No market source or evidence-consuming rate acceptance exists in this method; timestamp records now, not rate vintage.

Concrete deduction, not execution: numeric input 100 USD to EUR passes both table memberships; USD intermediate is 100/1.0, target is 100×0.92, returned conversion is 92.0. The result is entailed by the implementation and assumed inputs, not by knowledge of present market rates.

"USD": 1.0, "EUR": 0.92, --- chapter1/context/agent.py

usd_amount = amount / exchange_rates[from_currency] converted_amount = usd_amount * exchange_rates[to_currency] --- chapter1/context/agent.py

"converted_amount": round(converted_amount, 2), "exchange_rate": round(exchange_rates[to_currency] / exchange_rates[from_currency], 4), "timestamp": datetime.now().isoformat() --- chapter1/context/agent.py

EPI-RTE-grounding-check

Label: Numerical grounding check

Part of: RT-RTE-context-ablation

Endpoints/progression: sent API message snapshots → tool-message quantities → answer/task quantities → approximate matching → verdict and unsupported list. Owner: ablation driver and assess_groundedness. Target: revenue-scale figures extracted from the terminal answer, relative to task and delivered tool-message figures. Evaluator: regex/scaling extraction, floor 100,000, relative tolerance 0.001, and a fixed branch order. It does not compare propositions or reference task answers. Object: EPI-OBJ-final-answer. This subroute fills the evaluator implementation gap expressly left by the supplied ablation record, without replacing that record. SRC-1: chapter1/context/grounding.py, chapter1/context/main.py, implementation.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Verdict, observation count, extracted answer quantities and unsupported subset.
• Later read-back: inapplicable — current answer assessment does not modify retained agent content or evaluator state for later invocations.
• Delegated visibility: Local program consumes sent-message snapshots, not withheld local tool results; no provider call is made by this assessor.
• Selection predicate: Tool-role message quantities are extracted with floor zero; answer/task use the default floor. A nonempty observations list forces not_assessable after the no-answer branch.
• Invalidation or expiry: Recomputed per returned test; no retained evidence acceptance or withdrawal mechanism.
• Activation or effect: Check wiring is implemented; report consumption is separately EPI-RTE-grounding-report. The function itself returns evidence without a behavioral veto.
• Evidence limits: Matching establishes availability of a similar numeral, not actual use or truth; unmatched output can be valid arithmetic. Regex extraction is not semantic parsing, and smaller claims are ignored.

Operating mode: bounded case/mode comparison. Human supplies task/case/mode launch inputs; software computes verdict; no model judges or vetoes it. Answer oracle: none for task correctness; the software predicate is the oracle only for its own defined matching classification. This route does not admit a product or machinery update.

Counterexample to the stronger evidence claim, deduced without executing target code: task Add 100000 and 200000, answer 300000, observations []. Extraction returns task quantities 100000 and 200000, answer quantity 300000. Both relative differences exceed 0.001, so the answer is marked ungrounded, although the asserted sum follows from task evidence. Conversely, copying a task number can pass while asserting the wrong relation. Therefore the result is not a derivation-provenance or correctness oracle.

QUANTITY_FLOOR = 100_000.0 --- chapter1/context/grounding.py

DEFAULT_REL_TOL = 1e-3 --- chapter1/context/grounding.py

known = extract_quantities(task_text) + list(observations) unsupported = [q for q in quantities if not matches_any(q, known)] --- chapter1/context/grounding.py

elif observations: # With real observations in context there is no honest way to tell a # correct mental calculation from an invented number, so say so instead # of manufacturing a verdict. result["verdict"] = "not_assessable" elif not quantities: result["verdict"] = "no_quantities" elif not unsupported: result["verdict"] = "grounded" else: result["verdict"] = "ungrounded" --- chapter1/context/grounding.py

for turn in result["trajectory"].api_turns for message in (turn.get("request") or {}).get("messages", []) --- chapter1/context/main.py

EPI-RTE-grounding-report

Label: Grounding report consumption

Part of: RT-RTE-context-ablation

Endpoints/progression: grounding dictionary → test result → logger warning, analysis counts/failure reason and displayed table. Owner: AblationTestSuite. Target/condition: verdict equal to ungrounded, with table mapping for other verdicts. This is the report-consumption subroute, not the numerical check or candidate release. SRC-1: chapter1/context/main.py, implementation. Closest records: parent ablation route and EPI-RTE-grounding-check; distinct function and consumer force.

• implementation conclusion status: wired
• operation conclusion status: uninspected
• Immediate return: Summary fields, unsupported counts/modes and advisory failure reason; warnings/table output when called.
• Later read-back: inapplicable — inspected consumers do not reload the report into future agent or policy decisions.
• Delegated visibility: Operator sees report labels; the checked answer has already returned from the model.
• Selection predicate: Equality to ungrounded selects warning, counts and unsupported-number failure reason; printed rows use verdict-to-label mapping.
• Invalidation or expiry: Current report generation only; no implemented answer withdrawal or claim endorsement lifecycle.
• Activation or effect: Advisory reporting is wired. Consumer: logger/report builder/operator; channel: test-result fields and displayed table; force: advisory classification; horizon: this comparison report. No enforcing correctness gate or model repair follows.
• Evidence limits: No observed report inspected; report labels cannot enlarge the check's numeric-domain license or establish actual provenance.

Operating mode: bounded ablation reporting; program decides labels, human reads them; no per-content human acceptance, successor selection or task reference oracle.

if test_result["grounding_verdict"] == "ungrounded": logger.warning( "Answer states %d figure(s) with no observation behind them: %s", len(test_result["unsupported_quantities"]), test_result["unsupported_quantities"], ) --- chapter1/context/main.py

mode_analysis["failure_reason"] = ( "Answered with figures no tool observation supports" ) --- chapter1/context/main.py

Claims

EPI-CLM-current-rates

Label: Current exchange rates

Claimed operation: the tool schema says conversion uses current exchange rates. SRC-1: chapter1/context/agent.py, doctrine/design delivered to the model. No counterpart among supplied claims: RT-CLM-context-purpose concerns ablation intent, and RT-CLM-canary-decision concerns candidate eligibility. The runtime prose already notes this mismatch; this declaration isolates the consequential tool claim rather than duplicating a supplied claim record.

"description": "Convert an amount from one currency to another using current exchange rates", --- chapter1/context/agent.py

EPI-CLM-grounding-source

Label: Grounding source warrant

Claimed warrant: with no observations, figures not supplied by task text are classified as lacking evidence. SRC-1: chapter1/context/grounding.py, doctrine/design. Distinct from the supplied teaching-purpose and canary claims; closest is RT-CLM-context-purpose, which names purpose but not this provenance license. The implemented counterexample is in EPI-RTE-grounding-check. The source itself distinguishes grounding from correctness, but that distinction does not remedy the arithmetic-closure gap.

The model saw no observations yet stated figures the task never gave it. Whatever produced them, it was not evidence. --- chapter1/context/grounding.py